Using audio files, Mel Frequency Cepstral Coefficients (MFCCs), and obtaining models features, this paper proposes the two-stage method for emotion recognition in speech using deep learning's Bidirectional-Long Short-Term Memory (Bi-LSTM) approach.. The suggested novel feature extraction method is used to build a Bi-LSTM that sequentially processes input sequences in both directions. This bidirectional processing enables the network to acquire information from both the past and future contexts, making it well-suited for tasks that need a thorough grasp of sequential data that contributes to emotion recognition. Including anger, happiness, sadness, disgust, fear, and neutrality from audio signals. Training and test datasets are analysed in Python Colab to produce good results compared to other deep learning techniques. The CREMA-D and SAVEE datasets are used to recognize precision, sensitivity (true positive rates) of uttered emotions, and spectrograms of human emotions.